A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 12:15 UTC
[model] make llama4 compatible with pure dense layers (#17315)
Signed-off-by: Lucia Fang <fanglu@fb.com>
Lucia Fang committed
Apr 28, 2025 at 19:22 UTC
b4ac4fa04da14c4f40688fb98211367981a1f4d7
1 file changed
+2
-2
vllm/model_executor/models/llama4.py
+2
-2
@@ -273,8 +273,8 @@ class Llama4DecoderLayer(nn.Module):
273
cache_config=cache_config,
274
prefix=f"{prefix}.self_attn",
275
)
276
- is_moe_layer = (self.layer_idx +
277
- 1) % config.interleave_moe_layer_step == 0
276
+ is_moe_layer = config.interleave_moe_layer_step > 0 and (
277
+ self.layer_idx + 1) % config.interleave_moe_layer_step == 0
278
if is_moe_layer:
279
self.feed_forward = Llama4MoE(
280
config=config,